CITIC SEC: DeepSeek-V4.1-Flash released, focus on three investment mainlines

CITIC SEC released a research report stating that DeepSeek-V4.1-Flash has been released, achieving improvements in Coding and Agent capabilities while strengthening multimodal understanding. The model innovatively adopts the CED architecture, combining CSA2 cross-layer cache sharing, index reuse, and FP4 quantization, reducing the global KV Cache to approximately 1/4 of V4-Flash. Architecture optimization unlocks room for inference cost reduction, and DeepSeek announced further API price cuts. Domestic models will continue to compete along two paths: high cost-effectiveness and frontier intelligence. Under DeepSeek's deep adaptation, domestic computing infrastructure and domestic models will benefit in tandem, while high cost-effective inference will accelerate the deployment of FDE and enterprise-level Agents. CITIC SEC's main views are as follows: DeepSeek-V4.1-Flash: Adopts the new CED architecture, with model capabilities surpassing DeepSeek-V4-Pro. On September 10, 2026, DeepSeek released DeepSeek-V4.1-Flash, adopting a brand-new "Causal Encoder-Decoder" (CED) architecture, featuring multimodal understanding capabilities, supporting million-token context, and simultaneously opening API and model weights. The model backbone has 552 billion parameters, activating 8 billion and 16 billion parameters per Token during the prefill and decode stages respectively, further reducing inference overhead in long-context and Agent scenarios. According to DeepSeek's official WeChat public account, considering that V4.1-Flash has comprehensively surpassed V4-Pro in performance, cost, speed, total time, and other metrics, DeepSeek plans to gradually take V4 Pro offline. However, at this stage, considering that V4-Pro still maintains a large call volume and Infra migration requires time, as of September 15, V4-Pro remains in service. Coding and Agent capabilities significantly improved, multimodal understanding expands task execution boundaries. According to DeepSeek's WeChat public account, 1) In terms of Coding capability, at the highest thinking intensity, V4.1-Flash scored 90.6 on Terminal-Bench 2.1, higher than Opus 5's 89.1 and GPT-5.6 Sol's 88.8. In the more difficult Terminal-Bench 4.0, V4.1-Flash scored 31.2, lower than GPT-6 Astra's 57.9 and Fable 5.1's 55.8; in the DeepSWE v1.1 test, it reached 74.2, roughly on par with GPT-6 Astra's 74.1 and higher than Fable 5.1's 67.4. 2) In terms of Agent capability, AutomationBench and Agent's Last Exam scored 54.8 and 31.8 respectively, both higher than Opus-5's 50.3 and 28.6 and GPT-5.6 Sol's 45.8 and 26.7, reflecting that after enhanced visual understanding capabilities, the execution ability for general automation tasks and complex tasks has improved. Architecture innovation: CED architecture and CSA2 collaboratively optimize computation paths and cache reuse, continuing the ultimate cost reduction of long-context inference. 1) Innovative "Causal Encoder-Decoder" (CED) architecture, reducing global KV Cache usage to approximately 1/4 of V4-Flash. The model's 40-layer Transformer network is divided into the first 20 layers as Causal-Encoder and the last 20 layers as Decoder. The global KV required by the decoder is directly projected from the encoder's final layer output, without relying on the hidden states of each decoder layer to generate layer by layer, so most input tokens only need to complete encoder computation. To preserve local information, the decoder additionally processes the last 128 tokens of the prompt, approximately reconstructing the sliding window attention (SWA) cache of each layer, reducing the prefill computation in long-input scenarios by nearly half. According to DeepSeek's official technical report, the activated parameters per token during the Prefill stage decrease from 16 billion in the decode stage to 8 billion. In terms of cache precision, the model reduces the global main KV Cache from the previous generation's FP8 to FP4, and introduces quantization-aware training during post-training to control precision loss, reducing this part's storage usage by nearly half, while SWA KV retains FP8. The model architecture innovation, combined with CSA2's cross-layer cache sharing mechanism, reduces global KV Cache usage to approximately 1/4 of V4-Flash, simultaneously driving the new model's HBM demand down to 1/4 of V4-Flash and SSD demand down to 1/8. 2) CSA2 simultaneously reduces cache usage and floating-point computation through layered KV sharing and reuse of Top-K token retrieval results. The model's first two layers use only SWA, while the remaining layers adopt compressed sparse attention CSA2, configured with three modes: Full, Reindex, and Reuse. In Full mode, the model generates the global main KV and index keys through computation, and filters Top-K historical positions; Reindex shares the previous layer's KV from Full mode in the next layer, but recalculates relevance to update Top-K tokens; Reuse simultaneously shares KV and Top-K results, eliminating repeated cache generation and index computation. Specifically, the encoder's remaining 18 layers besides the 2 SWA layers are configured in groups of 6 layers as "1 Full layer + 5 Reuse layers"; the decoder's 20 layers are divided into 5 groups, with the first group being "1 Full layer + 3 Reuse layers" and the latter 4 groups being "1 Reindex layer + 3 Reuse layers," enabling the latter 20 layers to share the global KV generated by the first Full layer. The above design combines reducing KV Cache with reducing Indexer retrieval computation. According to DeepSeek's official technical paper, when the model context expands from 4K to 1M, single-Token decode FLOPs increase by only about 25%, demonstrating significant improvement in long-context inference efficiency. Inference optimization transmits to API price cuts, maintaining the high cost-effectiveness advantage of domestic open-source models. According to DeepSeek's official pricing, V4.1-Flash's per-million-Token off-peak cache-hit input, cache-miss input, and output prices are 0.02 yuan, 1 yuan, and 4 yuan respectively, down 60.0%, 33.3%, and 11.1% from the previous generation's 0.05 yuan, 1.5 yuan, and 4.5 yuan. Peak-hour prices are twice the off-peak prices. Application impact: The new model continues the ultimate cost reduction direction, benefiting FDE and enterprise-level Agent application deployment. While improving coding, Agent, and multimodal capabilities, the model continues to lower call prices, which is expected to reduce the cost burden of multi-turn tool calls and long-context processing for enterprise-level Agents, promoting accelerated FDE deployment. At the same time, professional software companies with industry knowledge accumulation, deep embedding in enterprise business processes, or vertical private-domain data barriers that can deliver reliable results in highly regulated scenarios are expected to be the first to transform model capabilities into product value and commercialization increments. Investment strategy: Recommend focusing on the following three investment mainlines. 1) AI infrastructure: DeepSeek deeply adapts to domestic computing power, with domestic computing power and domestic models moving toward each other. 2) AI applications: The model continues its open-source strategy, further improving inference efficiency and reducing costs, benefiting FDE and software application companies with barriers. 3) Model original manufacturers: On one hand, architecture innovation continues to unlock room for inference efficiency improvement and cost reduction. DeepSeek-V4.1-Flash, GLM-5.3-Flash, and Qwen3.8-Flash released within the past 2 months all significantly reduced long-context inference overhead through architecture optimization, verifying that architecture upgrades still possess further cost reduction potential. On the other hand, domestic models will continue to compete along two paths: high cost-effectiveness and frontier intelligence levels. The Flash series covers daily coding, office, and general Agent needs with lower activated parameter counts and call costs, while large-parameter flagship models are expected to continue raising the upper limits of complex reasoning, professional knowledge, and long-horizon task capabilities. Risk factors: AI core technology development and application expansion falling short of expectations; computing power cost reduction falling short of expectations; AI being improperly used causing severe social impact; data security risks; information security risks; intensifying industry competition, etc.